Learning Category Instances and Feature Utilities in a Feature-

Selection Model

Michael Martin and Christian Lebiere Carnegie Mellon University, Pittsburgh, PA, USA

MaryAnne Fields and Craig Lennon

Army Research Laboratory, Aberdeen, MD, USA Presented at the 2016 ACT-R Post-Graduate Summer School, Lancaster, PA

Autonomous Systems

• Perceptual systems tend to feed forward to cognitive systems that provide little feedback

• Establish a feedback loop between perceptual and cognitive systems

• Exploit cognitive context to augment bottom-up perceptual approaches

Autonomous Perception

Autonomous Cognition

Intent

• ACT-R = a priori architecture of ML methods • Traditionally, ML methods tailored to problem ad-hoc • Classification & Feature Selection usually treated separately

• Create generic cognitive module for classifying objects, scenes or situations

• Feature Selection Model (YACM) • FS Model = Prediction problem (which class) + Inference problem

(important features)

• Design goals • Take input from any perceptual module • Learn to classify (Instance-Based Learning) • Learn which features to use (Feature Selection) • Stay close to architecture for generalizability

• Respect default parameter values

• Observe strict harvesting

• Fixed representational scheme will not work • Flexible chunks in v7 make this easy

EGCM-RT (Lamberts, 2000) & EBRW (Nosofsky & Palmeri, 1997)

• Encode features one at a time until a category decision is made • Decision depends on similarity b/t S & instances in

memory

• S representation gradually constructed via info accumulation process

• Accumulation stops as function of evidence for category membership

• Retrieved exemplars during info accumulation drive a random walk process • Implies a random walk pointer

• Exemplars retrieved until pointer exceeds criterion value for 1 of classes

Feature Selection Model • Combine IBL w/FS to anchor perceptual labels

• S information accumulated in imaginal buffer

• Make class inference for each feature encoded

• Random walk pointer & criterion value represented by relative utility of & competition among encoding & decision productions

Declarative Memory IBL Prediction problem

Procedural Memory R-L Inference problem

Infer label given observed features

Encode feature given label

Classify exemplar

Utility

Cohn-Kinade Facial Expressions Surprise

Happy

Disgust

Sad

Anger

au34

au24

au23

anger au9

disgust

au17

au15

au26

sadness

au2

au39

fear au4

au7 contempt

au20

au14

au21

au11

au10 au1

au25

surprise happy

au12 au6

unknown

au5

au27

unknown

happy

surprise

sadness

disgust

anger

fear

contempt

Emotion x Feature network indicates that emotion labels are associated with both shared- and distinct facial Action Units.

Categorization depends on different subsets of overlapping and non-overlapping features.

Similarity of emotion labels based on shared AUs (features).

Baseline: Retrieve Exemplars

That's 20% correct

10,000 trials

Proportion correct for sliding window 100 trials wide

Activation :bll nil :mas 10 :ans nil

Utility Reward 10, -1 :alpha 0.2 :egs 0

Anger everywhere 188 chunks

Retrieve Exemplars w/Noise

Activation :bll nil :mas 10 :ans .25

Utility Reward 10, -1 :alpha 0.2 :egs 1.0

That's 25% correct

0 1

anger 4266 663

contempt 646 30

disgust 1457 323

fear 795 66

happy 1533 454

neutral 10 0

sadness 892 81

surprise 1684 584

unknown 16 0 3695 chunks

Noise improves response variability

Retrieve Situated Prototype

That's 50% correct

Utility Reward 10, -1 :alpha 0.2 :egs 1.0

Mix of exemplars & prototypes yields better performance

11,562 chunks

Augment Context Effect

Activation :bll nil :mas 10 :ans .25 :tmp 10

Utility Reward 10, -1 :alpha 0.2 :egs 1.0

That's 60% correct

11,591 chunks

Confusions After 9000 Trials

Overgeneralization Many false alarms

Control Reward Distribution

Activation :bll nil :mas 10 :ans .25 :tmp 10

Utility Reward 10, -1 :alpha 0.2 :egs 1.0

Add "guessing" productions for when inferred class changes just before conclusion is drawn.

Distribute NULL rewards when inferred class changes

9721 chunks

That's 80% correct

Confusions After 9000 Trials

Fewer false alarms

Class-specific Decisions

10,161 chunks

Activation :bll nil :mas 10 :ans .25 :tmp 10

Utility Reward 10, -1 :alpha 0.2 :egs 1.0

Add 1 decision & 1 "guessing" production for each class as a means for representing costs/benefits associated with particular decisions.

That's 80% correct

Confusions After 9000 Trials

Feature Counts by Class

df$response: anger Min. 1st Qu. Median Mean 3rd Qu. Max. 2.00 9.00 16.50 18.69 29.00 39.00 ------------------------------------------------------------ df$response: contempt Min. 1st Qu. Median Mean 3rd Qu. Max. 6.00 22.75 27.50 28.82 38.00 39.00 ------------------------------------------------------------ df$response: disgust Min. 1st Qu. Median Mean 3rd Qu. Max. 1.00 8.00 21.00 21.17 33.00 39.00 ------------------------------------------------------------ df$response: fear Min. 1st Qu. Median Mean 3rd Qu. Max. 6.00 18.00 28.00 25.89 35.00 39.00 ------------------------------------------------------------ df$response: happy Min. 1st Qu. Median Mean 3rd Qu. Max. 1.0 13.0 24.0 23.1 34.5 39.0 ------------------------------------------------------------ df$response: sadness Min. 1st Qu. Median Mean 3rd Qu. Max. 4.00 13.75 23.50 22.84 30.50 39.00 ------------------------------------------------------------ df$response: surprise Min. 1st Qu. Median Mean 3rd Qu. Max. 0.00 2.00 11.00 14.64 26.00 39.00

au34

au24

au23

anger au9

disgust

au17

au15

au26

sadness

au2

au39

fear au4

au7 contempt

au20

au14

au21

au11

au10 au1

au25

surprise happy

au12 au6

unknown

au5

au27

Data: Feature presence only

Model: Feature present or absent

Gold & Coal

• Gold Nugget • Flexible chunks

• Link b/t symbolic & subsymbolic representation

• Subsymbolic learning that is a mix of traditional ML methods

• Lump of Coal • Lack of a "simulated annealing" process in R-L

• Rewards tied to time scale

• Lack of production-specific exploitation/exploration